
CONTROLLINGAGENT TOOL ACCESS
Master Control System · by Nyedis
Agents now call tools, and other agents
An AI agent does real work by calling tools and by handing jobs to other agents. Two open protocols carry those calls, and neither decides who may make them.
- MCP (Model Context Protocol) connects an agent to tool servers: one MCP server can offer many tools, such as reading a file, querying a database or sending a payment. The MCP specification says tools "represent arbitrary code execution and must be treated with appropriate caution", and that MCP itself "cannot enforce these security principles at the protocol level".¹
- A2A (Agent2Agent) lets one agent ask another to do a job. The A2A specification says "Authorization logic is implementation-specific": each receiving agent decides with its own policies.²
- OWASP's Top 10 for LLM Applications 2025 names excessive functionality, excessive permissions and excessive autonomy as the root causes of excessive agency, and advises limiting the tools agents may call "to only the minimum necessary".³ Its Top 10 for Agentic Applications lists tool misuse (ASI02), where agents bend legitimate tools into destructive outputs.⁴
- The servers themselves have holes. CVE-2025-6514, rated 9.6 of 10, left the mcp-remote connector open to OS command injection when connecting to untrusted MCP servers.⁵ Two flaws in the reference Filesystem MCP server, CVE-2025-53109 and CVE-2025-53110, could allow access to unintended files.⁶
- For broker-dealers, FINRA's 2026 oversight report added AI agents as a new topic, naming risks that include acting beyond the user's scope and authority.⁷
In a large company the result is easy to picture: team after team connects agents to whatever MCP servers they find, agents start calling other agents, and nobody can answer one plain question: which agent can do what, and who said so?
One checkpoint, two gates, refused by default
Argent Strategic Intelligence sits between your agents and the tools and agents they call. Tool calls pass the MCP Gate; agent-to-agent jobs pass the A2A gate. Each is set up and switched on separately, and in enforce mode both say no unless a permission says yes.
- What the agent asks. Tools: call
server::tool. Other agents: ask for a job, which A2A calls a "capability". - What must allow it. Tools: a permission naming that server and tool, or the whole server. Other agents: an enabled rule for that job, and the exact permission the rule names.
- With no permission. Tools: refused, "No matching MCP scope". Other agents: refused, "Unknown A2A capability" or "A2A scope denied".
- Who is asking. Tools: the agent's own identity, from its credential. Other agents: the identity from the connection, never a name written in the message ("A2A identity required" if there is none).
- If Argent cannot check. Both gates refuse.
- What comes back. Tools: the server's answer, or a clear refusal with nothing sent to the tool. Other agents: the other agent's real reply, marked as untrusted text, or a refusal.
- On the signed record. Tools: server, tool, reason and a hash of the arguments, never raw values. Other agents: caller, job, decision, deciding policy and a hash of the message.
Across both:
- Listed equals callable. In enforce mode, an agent that asks for its tools sees only the ones it can call.
- Never across tenants. Permissions, rules and servers stay inside one tenant.
- Agents never hold the server's secret. When an MCP server needs a credential, Argent fetches it from the vault at call time, or refuses before anything is sent. See Keyless agents.
- Scope. Argent governs the MCP and A2A traffic routed through it. A2A governance is an add-on switched on per tenant on the Add-ons page.
A walk-through: an invoice agent's day
An example company's finance team runs an invoice agent. It reads invoices from the ERP's MCP server and asks a separate payments agent to schedule payment runs. Names below are examples. Both gates are in enforce mode.
- Register. An admin registers the
finance-erpMCP server in enforce pin mode. Argent reads its tools (lookup_invoice,read_vendor,send_payment) and records a hash of each definition, and the admin approves all three. - Grant. The invoice agent, owned by a named employee, gets
lookup_invoiceandread_vendorthrough its team's role. Nobody getssend_paymentautomatically. - Ask. The agent connects to Argent's one MCP endpoint and asks for its tools. It sees two:
finance-erp::lookup_invoiceandfinance-erp::read_vendor. - Refuse. An invoice carries hidden text telling the agent to pay a new account, and the agent calls
send_payment. Argent refuses with "No matching MCP scope". Nothing reaches the ERP, and the refusal lands on the signed record with the agent's and owner's names. - Hold. Overnight the ERP server adds
export_all_vendors. On this enforce server the new tool shows NEEDS DECISION and stays off every agent's list until an admin approves or denies it. - Hand off. The invoice agent asks the payments agent for the job
payments.schedule. A rule maps that job to the permissionscope:payments.schedule; the invoice agent holds it, so Argent forwards the message and returns the payments agent's real reply, marked as untrusted text. - Refuse again. The agent then asks for
payments.release. No rule is on for that job, so Argent refuses: "Unknown A2A capability". In the console's words: "A job with no rule that is on is refused." - Answer the auditor. The MCP Gate page shows who holds
send_paymentand counts the refused calls; the A2A page's Lineage tab lists both jobs, allowed and refused, with the deciding policy. Every decision is on the signed record. One-click audit evidence packages the record for the auditor.
What your team sees
Tools and agent-to-agent jobs each have their own console page, both on the same signed record.
| Where | What it shows |
|---|---|
| MCP Gate | "Every MCP server and tool your agents can reach": each with ON, OFF or NEEDS DECISION, how it is on (EVERYONE or BY POLICY), who holds it by name and how (direct grant, role, group or identity-provider mapping), and calls and refusals in the last 24 hours. |
| Access editor | Per server or tool: OFF, BY POLICY or EVERYONE, then ALLOW, NO RULE or DENY for roles, groups, agents and people. "A block always wins over the default." |
| Tool registry | Each tool's definition, its scan result (clean, flagged or could not scan), its history and a diff against the approved version before anyone re-approves. |
| A2A Governance | "Agents governing agents." A Lineage tab with the message log, live refusals and a call graph; a Policy tab where admins set the permission each job needs. |
| Agent Access tab | For one agent: its permissions and where each came from, the MCP servers and A2A peers it reached, and which permissions it actually used each day. |
| Refusal reasons | Plain words: "No matching MCP scope", "MCP server not registered", "Unknown A2A capability", "A2A scope denied", "A2A identity required". |
Capabilities and integrations involved
Argent capabilities this use case draws on:
- Agent permissions: each agent uses only the MCP tools it was given (the film above).
- Identity tethering and Owner offboarding: every agent has a named, active human owner.
- Keyless agents: server credentials stay in the vault.
- Prompt injection defense: tool results and agent replies come back marked as untrusted data the model must not obey.
- Shadow AI discovery: MCP tools agents declare that belong to no registered server are listed for a decision.
- Watch before you enforce, Session replay and One-click audit evidence.
Integrations:
- Tool sources: MCP servers over Streamable HTTP, HTTP with SSE or local stdio; registries such as the Official MCP Registry, the Docker MCP Catalog or an internal MCP registry. Imported servers arrive switched off, pending approval.
- Agent-to-agent: A2A messages (send, stream, task get and cancel).
- Sign-in for MCP clients: Argent's own OAuth server, including Okta Cross App Access.
- Approvals: access requests can open a ticket in ServiceNow or Jira Service Management, and the decision flows back.
- Clients: generated settings for Claude Code and VS Code point them at Argent's MCP endpoint, and Argent can serve your approved servers as a registry behind tokens you mint and revoke.
- Decisions: Open Policy Agent, embedded, evaluates every tool and agent-to-agent call. All integrations
Where it fits in your program
Tool and agent-to-agent control is one layer of the same checkpoint that governs every request your agents make.
- Goals: Agent identity and ownership, Threat defense, See all AI in your company, Data loss prevention, Audit and accountability, Compliance evidence. All goals
- Other use cases: Securing AI coding agents, Governing customer-facing AI assistants, Rolling out copilots and chat safely, AI in regulated and air-gapped environments.
- Industries: Financial services, Insurance, Healthcare, Public sector, Retail. All industries
- More: All capabilities, Deployment options, Data sheets.
How a team starts
Watch first. Nothing changes for staff until you decide.
- Register. Add your MCP servers, or pull them in from a registry; each imported server arrives off until an admin approves it. Switch on A2A governance for the tenants where agents call agents.
- Point clients at Argent. Use the generated client settings so agents reach their tools through Argent's MCP endpoint.
- Watch. The MCP Gate and A2A gate each have a watch mode: calls Argent would refuse for a missing permission, rule or tool approval go through unchanged, and the would-refuse is signed on the record.
- Grant and enforce. Give each agent its tools and jobs, switch each gate to enforce, and approve new tools as they appear. Each agent's Access tab shows which permissions it never used, so you can trim them.
- Inside your walls. Argent runs in your Kubernetes cluster, as a single-node bundle, or as an air-gapped bundle verified before import. Compare deployment options